The Infrastructure Squeeze
Navigating supply chain bottlenecks to accelerate AI deployment
Don’t assume that having the capital means you can build the AI infrastructure you need. Hyperscalers are consuming the global manufacturing pipeline for GPUs, CPUs, memory, disk drives, and skilled construction labor, and enterprises competing for the same equipment are finding that money doesn’t move them up a 2-year queue.
Lead times are worse than most leaders realize. In my experience, even with a non-cancelable order, lead times for GPUs, CPUs, and other server components are now well past a year. This is even worse for components required for the buildout that are not currently manufactured in high volume. SAVRN’s 2026 supply chain analysis shows high-horsepower gensets are sold out through 2028, and Reuters reports that step-up transformer lead times now exceed 160 weeks, driven by data-center growth.
Construction labor is another factor that leaders do not adequately account for. DC Atlas reports that contractors are walking away from projects due to staffing shortages, and grid connection is taking longer than 5 years. Industry groups warn that workforce gaps in transmission, substation construction, and data-center trades will persist through the decade.
Even the hyperscalers are running into supply constraints. AWS, Azure, and Google Cloud are still growing fast, but they’re also projecting capacity constraints into the late 2020s.
I think “should we build our own AI cluster” is not the right question anymore. The lead times make it moot. The real question is how you accelerate your AI deployments without waiting for an infrastructure build that can keep you in the supply chain line for 2 years.
1. Buy time with hyperscaler capacity
Rent reserved cloud capacity while you secure long-lead equipment in parallel. Model development, data pipeline buildout, and agent deployment don’t need to wait 24 to 36 months for a transformer to arrive.
However, don’t skip the critical architectural requirement: keep proprietary workloads portable from day one. Portability must also account for data migration costs, or you will incur a large data egress bill when you are ready to move back on-prem. Without built-in portability, the temporary rental becomes permanent, and switching later costs far more than building it in from the start.
2. Secure long-lead items before you pick a site
Hyperscalers reserve equipment first and decide sites later. Enterprises considering datacenter buildouts must run the same play: sign multi-year agreements for long-lead items before starting the site search.
The most important thing I tell any leader weighing a full datacenter build is to engage with utility companies before selecting a site. Even if you engage with the utilities, you may still face uncertainties after site selection. This week, the Governor of Texas asked all datacenter projects to be paused until a full power grid audit is completed. Site connectivity also deserves special attention. I’ve seen connectivity delays surface 18 months into construction, delaying the bringing of a site online.
The Shift to Watch
AI infrastructure buildout is flush with capital. Don’t tie AI deployments to new sovereign enterprise infrastructure. Enterprises can’t outspend hyperscalers for scarce equipment, but they must maintain AI deployment velocity. Put strategies in place to maintain velocity by renting capacity and staging the infrastructure buildout to align with supply chain constraints.